Feat session preprocessing structure

[Creylay]

Summary

Redesigns the preprocessing step of the model session wizard so users build converter chains knowing, at every step, which columns exist and what type they have.

Before this change the wizard asked for converters first and inputs/output last, and the frontend guessed each converter's output type on its own. It never removed columns a converter consumed (after PCA on age, weight, both were still offered) nor updated columns replaced in place, so a chain could look valid and only fail when the session trained.

Now:

• The user picks the output column and the candidate columns first, then builds the chain, then picks the model inputs from what the chain leaves. The output column is never part of any converter scope.
• Every converter declares what it does to its columns (COLUMN_OPERATION: replace, add, expand, select, rows) and can describe its exact output with infer_output_columns.
• A new backend function, infer_structure, estimates the dataset state after every step without fitting anything. The wizard calls it on every change; session creation uses the same function and rejects invalid chains with a 422 before creating anything.
• Output columns whose names are only known after fit (e.g. one-hot columns) are shown as symbolic blocks with their size, or "N" when unknown.

---

Type of Change

Check all that apply like this [x]:

• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

---

Changes (by file)

Structure estimation (backend)

• DashAI/back/preprocessing/structure_types.py (new): state items (ColumnItem, BlockItem), StructureDelta, StepStructure, StructureResult and the type_fields helper.
• DashAI/back/preprocessing/structure.py (new): infer_structure. Walks the chain resolving each scope against the current state, checks types and cardinality like the frontend does, applies each converter's delta with the same column renaming as the runtime, and marks steps after an invalid one as blocked. Also resolve_state_refs and StructureError.
• DashAI/back/converters/base_converter.py: COLUMN_OPERATION attribute, default infer_output_columns per operation, and column_operation in the metadata.
• DashAI/back/converters/structure_mixins.py (new): ComponentsOutputMixin, whose output size follows n_components (PCA family and kernel approximations).
• DashAI/back/converters/category/*.py: each category declares its COLUMN_OPERATION.
• DashAI/back/converters/** (sklearn, simple, Hugging Face and SAM3 converters): exact infer_output_columns overrides where the output follows from params, and operation overrides where a converter differs from its category.

Deterministic output types

A converter's output type now follows from its params and input types, never from the values, so the estimate always matches the real output:

• scikit_learn/simple_imputer.py: mean/median always produce Float.
• simple_converters/column_arithmetic.py, numeric_expansion.py: integer results stay Integer even with nulls (computed with pyarrow.compute).
• simple_converters/character_replacer.py: keeps its input type instead of turning all-digit text into Integer, a decision that was made per batch (also affects notebooks).
• simple_converters/type_cast.py: on_error="skip" adds a cast_may_skip warning.

Runtime fixes

• DashAI/back/preprocessing/session_preprocessor.py: supervised converters now receive the output column as y (they previously got none and failed); a new column renamed on a name clash is recorded under its final name.
• DashAI/back/preprocessing/column_ref.py: GroupColumnRef gains an optional name to reference one generated column whose name is known before fit (e.g. date_month). No migration needed.
• DashAI/back/converters/dataset_columns.py: the clash renaming moves to plan_new_column_names, shared by the runtime and the estimate.
• DashAI/back/job/preprocessing_job.py: passes the output columns to the preprocessor.

API

• DashAI/back/api/api_v1/endpoints/model_sessions.py:
  • New POST /model-session/preprocessing/structure (read only).
  • Session creation validates the chain and the input refs and returns a 422 with the details.
  • /validation types every input ref from the estimated structure; the old frontend computed converter_output_types is removed.
• DashAI/back/api/api_v1/schemas/model_sessions_params.py: PreprocessingStructureParams; ColumnsValidationParams takes preprocessing.

Frontend

• components/models/CreateSessionSteps.jsx: with preprocessing on, the order is prepare, base columns, preprocessing, final inputs. Without it the wizard is unchanged. Preprocessing steps are no longer sent if the switch was turned off after adding them.
• components/models/modelSession/BaseColumnsStep.jsx (new): output and candidate columns.
• components/models/modelSession/usePreprocessingStructure.js (new): debounced call to the structure endpoint that discards stale responses.
• components/models/modelSession/PreprocessingStep.jsx, SessionConvertersRightBar.jsx, ScopeStepSessionConverter.jsx, FormSessionConverterSection.jsx: the catalog and the scope picker only offer what exists at the end of the chain; row-changing converters are hidden.
• components/models/modelSession/AppliedConvertersView.jsx:
  • Each card shows what its step produces.
  • Invalid steps are outlined in red with a translated reason, and later steps are dimmed.
  • A new "Dataset after preprocessing" panel lists the final state.
• components/models/modelSession/SelectColumnsStep.jsx: final inputs come from the estimated final state (all selected by default); the output is fixed.
• components/models/modelSession/sessionColumnRefs.js: keys for named refs, itemToRef, stateToOptions, labelForRef; the frontend type inference (buildColumnKeysAndTypes, resolveDeclaredOutputSlots) is removed.
• components/models/modelSession/structureMessages.js (new): turns backend structure messages into translated text.
• components/models/modelSession/DivideDatasetColumns.jsx: optional inputLabel and outputDisabled.
• components/models/SessionInfoContent.jsx: scope labels via labelForRef.
• components/notebooks/converterCreation/ParameterStepConverter.jsx: the data leakage warning only shows in notebooks, since session preprocessing already fits on the training split.
• api/modelSession.ts, types/modelSession.ts: structure endpoint client and types.
• utils/i18n/locales/*/models.json: new strings in the 5 languages.

Tests

• tests/back/preprocessing/test_structure_matches_runtime.py (new): for every registered converter, compares the estimated structure with a real fit and transform; also fails if a new converter has no case.
• tests/back/preprocessing/test_structure.py (new), tests/back/converters/test_infer_output_columns_defaults.py (new), tests/back/converters/test_download_converters_structure.py (new).
• tests/back/preprocessing/test_session_preprocessor.py, test_column_ref.py, tests/back/api/test_model_session_api.py, test_run_preprocessing_guard.py: updated and extended.
• Frontend: sessionColumnRefs.test.js, AppliedConvertersView.test.jsx, usePreprocessingStructure.test.js (new).

---

Testing (optional)

1. Create a model session with "Apply preprocessing (advanced)" on and pick the output and candidate columns.
2. Add PCA with n_components=2 over two numeric columns: both disappear from the "Dataset after preprocessing" panel and are no longer offered to later converters.
3. Create a session.